HN user

shauntrennery

14 karma

Developer / Creator / Founder with interests in family, development, JavaScript, startups, news, location, fitness and all things web.

Posts6
Comments18
View on HN

Author here. I built Numble as a daily number-guessing puzzle in the Wordle mould, and along the way noticed that its feedback mechanic (coarse distance buckets with no direction) is unusually hostile to LLMs trying to play it cold. The article explains why: the answer never appears in the DOM, the feedback forces interval arithmetic across five independent candidate sets, and a recently added "joker" tile hides one cell's feedback entirely — which turns out to be the single most adversarial-to-LLMs feature in the game, and I shipped it as a visual gag. Happy to be told I'm wrong about the "first" framing; that's the part I'm least sure of.

The "close/warm/far" distance feedback is meant to make each column solvable in ~3 moves. 4 max with a bad start feels about right.

Only 16% of plays are solved in 3 or fewer guesses.

Curious if the all-ones strategy "feels" fun or if it feels like grinding.

Wordle, but for numbers. Guess the 5-digit code in 6 tries. Each guess reveals how close you are - but not which direction. New puzzle daily. Can you crack today's Numble?

#Do cards are similar to Google Now cards. The idea is to enable any small business to expose a digital interface. i.e. a small collection of single-function cards which can then be dispersed, surfaced and executed anywhere.

I definitely see them used in notifications. Another use case that really excites me is Google's eddystone standard.